NVIDIA H200 141GB SXM

ภาพรวม

The H200 extends the Hopper generation with 141 GB of HBM3e and 4.8 TB/s of memory bandwidth, so larger weights and longer context windows stay resident on one accelerator. It suits large language model training, fine-tuning and inference serving, plus HPC codes limited by bandwidth rather than raw compute. Note that the SXM5 module is not a card: it mounts on an HGX H200 baseboard, normally in groups of four or eight.

AXISLINK sources H200 modules and HGX baseboards through verified channels, and every serial number is documented on the quotation before you commit. We supply baseboards on their own or integrated into validated server platforms with power, cooling and networking specified for the deployment. Export documentation, insurance and freight are arranged from Hong Kong, and end-user information may be required before shipment for this class of accelerator.